Algorithm for Skeleton Action Recognition by Integrating Attention Mechanism and Convolutional Neural Networks

An action recognition model based on 3D skeleton data may experience a decrease in recognition accuracy when facing complex backgrounds, and it is easy to overlook the local connection between dynamic gradient information and dynamic actions, resulting in a decrease in the fault tolerance of the constructed model. To achieve accurate and fast capture of human skeletal movements, a directed graph convolutional network recognition model that integrates attention mechanism and convolutional neural network is proposed. By combining spacetime converter and central differential graph convolution, a corresponding central differential converter graph convolutional network model is constructed to obtain dynamic gradient information in actions and calculate local connections between dynamic actions. The research outcomes express that the cross-target benchmark recognition rate of the directed graph convolutional network recognition model is 92.3%, and the cross-view benchmark recognition rate is 97.3%. The accuracy of Top-1 is 37.6%, and the accuracy of Top-5 is 60.5%. The cross-target recognition rate of the central differential converter graph convolutional network model is 92.9%, and the cross-view benchmark recognition rate is 97.5%. Undercross-target and cross-view benchmarks, the average recognition accuracy for similar actions is 81.3% and 88.9%, respectively. The accuracy of the entire action recognition model in single-person multi-person action recognition experiments is 95.0%. The outcomes denote that the model constructed by the research institute has higher recognition rate and more stable performance compared to existing neural network recognition models, and has certain research value.

@article{Liu2023,
  title     = {Algorithm for Skeleton Action Recognition by Integrating Attention Mechanism and Convolutional Neural Networks},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {14},
  number    = {8},
  year      = {2023},
  publisher = {The Science and Information Organization},
  author    = {Jianhua Liu},
  doi       = {10.14569/IJACSA.2023.0140867},
  url       = {https://doi.org/10.14569/IJACSA.2023.0140867}
}